Methodology
Total latency equals TTFT plus output tokens divided by generation throughput. Approximate capacity uses concurrency divided by total request time.
Performance and Operations
Compare time to first token, generation throughput, output length, and end-to-end latency for two measured routes.
Interactive Tool
| Result | Value | How to read it |
|---|---|---|
| Route A total latency | 7.94 sec | 0.8s TTFT plus generation time. |
| Route B total latency | 5.67 sec | 1.5s TTFT plus generation time. |
| Completion-time difference | 2.28 sec | Route B completes sooner. |
| Route A capacity | 151.08 RPM | Approximate at entered concurrency and request time. |
| Route B capacity | 127.06 RPM | Approximate at entered concurrency and request time. |
Use measurements from the same model, provider, region, service tier, workload, output length, load, and percentile. The page does not supply or invent latency observations.
Inputs
| Input | How it is used |
|---|---|
| TTFT | Measured time to first token for each route. |
| Throughput | Measured output tokens per second after the first token. |
| Output length | Comparable generated tokens per request. |
| Parallel capacity | Concurrent requests available on each route. |
Total latency equals TTFT plus output tokens divided by generation throughput. Approximate capacity uses concurrency divided by total request time.
Use p50 and p95 measurements separately and keep model, provider, region, workload, tier, and timestamp fixed. This page does not supply unverified latency data.
Boundaries
Catalog values retain their source and freshness on the linked model, provider, benchmark, or comparison page. Editable scenario assumptions are not Model Markets measurements.
Continue the Analysis
| Tool | Next question |
|---|---|
| Rate Limit Calculator | Size an AI API quota from peak behavior rather than monthly averages. |
| GPU Inference Cost | Normalize a GPU deployment into cost per million generated tokens. |
| Model Router | Split routine and difficult requests across eligible models without pretending one model is optimal for every call. |
Questions
Normalize latency measurements into user-visible wait time for the same output workload. It returns time to first token, generation time, total latency, speedup, and requests-per-minute capacity for routes A and B.
Where the calculation needs model facts, it uses the current Model Markets catalog snapshot updated 2026-09-02. User-entered assumptions remain clearly editable, and unsupported values stay unknown rather than being inferred.
Throughput changes with sequence length, batch size, and load. Streaming improves perceived latency without changing completion time. Network and application overhead must be included in end-to-end measurements. Open the linked canonical records and primary sources before making a production or purchasing decision.